TLDR FOUNDERS 2026-10-23

Terms and Conditions apply, linked HERE [6].

THE CURRENT STATE OF AI TOKENOMICS | WHAT CFOS SHOULD KNOW

OpenRouter reports that open-weight models account for about 60% of weekly tokens on its platform but only 26% of spending. The price on the model list doesn't tell you what a completed task will cost. Cache discounts change the bill, and failed requests still cost money. If requests cost $1 and succeed 80% of the time, each successful result costs $1.25. Budget against successful tasks using your actual cache discounts and retry costs, then compare providers on that basis.

THE END OF STARTUPS

Most startups will not be able to buy compute within the next six months. It's becoming increasingly difficult to get capacity online for companies that are not investment grade. Startups are being asked to pay larger and larger upfront payments, and the few investment-grade tenants are getting increasingly down rated as they take on more and more debt and turn around to sell to buyers of tokens who themselves are not investment grade. Investment-grade buyers of GPUs will likely increasingly ask that the buyers of tokens be either investment grade themselves or pay larger deposits upfront.

HOW HARVEY IS BUILDING A FRONTIER MARKETING ORG

Harvey wanted a tailored demo film for hundreds of sales accounts, but each early attempt needed 15 to 20 minutes of back-and-forth. It first built reusable product screens and a library of animations in Figma. Agents then chose from 44 demo flows and used sales-call context to write each account's script. Harvey reports making 36 films in 69 minutes. Every film had a row in a shared tracker, and the rendering system produced contact sheets to flag clipped text and stray pointers before anyone watched.

HOW I USE SKILLS FOR EVERYTHING

A useful AI conversation can leave the next teammate rebuilding the same context. A PostHog engineer keeps shared skills that agents update at the end of each session. Unfinished work goes into a handover that gets overwritten as the work changes. Facts that will still be useful in a month get their own reference file. Before going on holiday, he linked teammates directly to those skills. New investigations also update the runbook, giving the next person the steps that resolved the issue.

INTRODUCING 1PASSWORD FOR AGENTS

The 1Password for Agents platform helps agents work while keeping the credentials they need secure and making access something every user can see, approve, and manage. Users can now ask agents to handle more useful tasks, as they can log into pages to complete tasks on their behalf. The user has control over what agents access and when throughout the whole process. 1Password for Agents is currently available at no additional cost for Individual and Family accounts on 1Password.com but not for accounts on 1Password.eu or 1Password.ca.

AUTOPILOT: ALWAYS-ON AGENT FOR SELF-UPDATING DOCS

Mintlify's Autopilot beta is available on Pro and Enterprise plans, using GitHub and Slack to draft documentation updates for human review at 250 credits per pull request opened or updated.

CLAUDE DUET

Delegate tasks to a cloud agent that works independently and reports back.

PRODUCT QUALITY

An AI feature can get worse without a new release. A model update or outdated source information can change its answers, and customers can start double-checking results or returning to spreadsheets without reporting a bug. Define acceptable performance using inputs that reflect actual customer situations, then keep evaluating after launch. Give someone responsibility for spotting deteriorating answers and turning each failure into a test the product has to pass.

HOW TO USE METRICS IN A FUNDRAISE

A healthy average growth rate can hide months when the business shrank. Before a fundraise, be ready to explain those swings and the assumptions behind each number. If you count active users by monthly logins, know what happens when the window shrinks to a day. The metrics you lead with should address the biggest doubt about your plan. For a startup counting on enterprise customers, that means showing it can win an enterprise sale.

HOW AN ANTHROPIC BDR IN JAPAN RUNS HIS DAY WITH CLAUDE MANAGED AGENTS

An Anthropic sales rep uses a shared knowledge base so he can correct an email once and have every agent reuse the fix. First replies and meeting follow-ups draw on the same saved edits. Each agent has one job, and only the rep can request changes to its instructions through a single Slack channel. Customer emails arrive as drafts for him to review and send. Routine CRM updates run on a cheaper model, while tasks needing more judgment use a larger one.

HOW TO BUILD AN AI BRAIN FOR YOUR POSITIONING

eWebinar put six years of customer evidence into a shared Claude project, tested its new positioning with conference prospects, then gave the team the same language guide to update the remaining copy.

YOU COULD DO IT ALONE

A cofounder can keep building while the CEO raises money, covering the work that still needs to happen when a capable solo founder's attention is elsewhere.

SWITCHING COSTS: HOW LLMS ARE THINKING ABOUT THEM. LESSONS FROM CHATGPT AND CLAUDE

Making conversation history easy to import gives users a way to try another product with their existing context.

BRINGING THE DTC ADS PLAYBOOK TO B2B

To get more distinct ads from your creators, give each one an audience or format your current ads don't cover.